2024-02-02-ENISA-2023年AI和标准化网络安全报告_37页_589kb

Cybersecurity of AI and Standardisation Report Summary

Introduction

This report provides an overview of existing, being drafted, under consideration, and planned standards related to the cybersecurity of artificial intelligence (AI), assessing their coverage and identifying gaps. It takes a broad view of AI, focusing primarily on machine learning (ML) and considers both the traditional CIA paradigm (confidentiality, integrity, availability) and broader trustworthiness aspects such as robustness and explainability. The analysis covers standardisation efforts by key organisations like CEN-CENELEC, ETSI, and ISO-IEC, and addresses potential alignment with the EU draft AI Act.

Definitions

• AI: Defined as software developed using techniques like machine learning, logic-based, or statistical approaches to generate outputs such as predictions or decisions, with a focus on ML.
• Cybersecurity of AI: Encompasses protection against attacks on AI systems' confidentiality, integrity, and availability, and considers a broader scope that includes trustworthiness features like data quality, transparency, and robustness to support overall security.

Standardisation in Support

Key standardisation organisations are actively developing standards:

• CEN-CENELEC: Focuses on transposing ISO-IEC standards to Europe; addresses AI through JTC 13 (cybersecurity) and JTC 21 (trustworthiness), proposing work on AI management systems and risk catalogues.
• ETSI: Via ISG SAI, producing group reports on AI threats, data security, and defence strategies; plans future standards for AI readiness, testing, and certification.
• ISO-IEC: Working through JTC 1 SC 42 on AI-specific standards, including risk management and data quality; however, standardisation needs are not fully harmonised.
• Gaps: Limited coverage of traceability, lineage, metrics for AI, and AI-specific attacks; overlap between AI and cybersecurity standards risks inefficiency.

Analysis of Coverage

• Narrow Scope (CIA): Existing standards like ISO/IEC 27001 can mitigate risks but require context-specific application; gaps exist in handling AI-specific threats.
• Broad Scope (Trustworthiness): Cybersecurity is intertwined with features like robustness and explainability; identification of incongruities in standards and areas for cohesion, such as in conformity assessment.
• AI Act Relation: Cybersecurity requirements apply to high-risk AI systems; standardisation gaps noted in competences for conformity assessment and keeping pace with R&D advancements.

Relation to Draft AI Act

The draft AI Act mandates cybersecurity for high-risk systems via a risk-based approach; standardisation supports implementation by providing technical requirements but faces gaps in comprehensive tools and evaluation methods. Recommendations include alignment with the Cybersecurity Act and fostering research to address evolving AI technologies.

Recommendations

• Organisations: Adopt harmonised AI terminology, develop guidance for applying existing security standards to AI.
• SDOs: Reflect ML-specific features in standards, establish liaisons for trustworthiness and cybersecurity integration.
• AI Act Prep: Promote sector-specific standards, fund R&D, standardise tools for auditing/certification, ensure legislative coherence.

Conclusion

Standardisation efforts in AI cybersecurity are advancing but face significant gaps, particularly in traceability and AI-specific metrics. A coordinated approach is needed to support the draft AI Act effectively, leveraging existing frameworks while addressing technological uncertainties through research and harmonised practices.